Capability records, perspective 3: Governance, risk, and accountability

Part of the SEI AI Adoption Framework (AIAF).

3.1 Acceptable-use policy

Purpose

Prevent people from guessing which AI uses are allowed, leading either to risky use or to blanket avoidance of useful tools.

Intended outcome

People know which AI uses are permitted, restricted, or prohibited and how to obtain guidance.

Decisions enabled

  • which uses are permitted, restricted, or prohibited;
  • what a person should do when a case is unclear;
  • when an exception may be granted; and
  • how breaches are handled.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • write the policy in plain language with examples of permitted and prohibited use;
  • cover data, tools, outputs, disclosure, and responsibility for results;
  • give people a way to ask and to request an exception;
  • obtain acknowledgement and refresh it when the policy changes;
  • apply the policy consistently, including to leaders; and
  • review the policy against questions, breaches, and new tools.

Required evidence

  • policy and examples;
  • acknowledgements;
  • exception records;
  • breach handling records; and
  • review history.

Common failure patterns

  • a policy too abstract to apply;
  • policy that bans tools people need, driving use out of sight;
  • no route to ask questions;
  • rules that differ across teams without reason; and
  • policy never updated.

Capability levels

Level Observable condition
Emerging Guidance is scattered in emails and informal advice, and people interpret it differently.
Defined A written policy with examples exists, and people acknowledge it.
Managed People can get guidance quickly. Exceptions and breaches are recorded and drive revisions.
Adaptive The policy adapts to new tools and observed behavior, and the organization tests whether people understand and follow it.

Review and exit triggers

  • a new tool, data type, or authority level appears;
  • breaches or repeated questions show the policy is unclear;
  • law, contract, or risk appetite changes; or
  • exception requests become routine.

Boundary

Tells people which uses are permitted, restricted, or prohibited. 1.7 sets what is unacceptable, and 3.3 sets how review depth scales.

3.2 AI inventory and classification

Purpose

Prevent the organization from relying on AI it cannot see, so that owners, data, connections, and lifecycle states are unknown when a risk or change arises.

Intended outcome

The organization can identify the AI solutions, features, tools, substrates, agents, owners, data, connections, and lifecycle states it relies upon.

Decisions enabled

  • what exists and who owns it;
  • which items need review or action;
  • what a change or incident would affect; and
  • which items to retire.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • define what counts as an AI solution, feature, tool, substrate, or agent;
  • record owner, purpose, workload, delivery mode, data, connections, and lifecycle state;
  • capture embedded features and tools used under personal or team accounts where possible;
  • record relationships between items, such as agents using tools and tools using data;
  • reconcile the inventory with purchasing, identity, and discovery records; and
  • review entries on a schedule.

Required evidence

  • current inventory with owners;
  • relationships and dependencies;
  • lifecycle status;
  • reconciliation results; and
  • review records.

Common failure patterns

  • an inventory of projects only;
  • entries that no one maintains;
  • embedded features missing;
  • no link between agents, tools, and data; and
  • inventory kept for audit rather than for decisions.

Capability levels

Level Observable condition
Emerging Knowledge of AI in use sits with individuals, and lists are incomplete or outdated.
Defined A maintained inventory records owner, purpose, data, and status for in-scope items.
Managed The inventory shows relationships, is reconciled with other records, and supports impact analysis and review.
Adaptive The inventory is updated from operating signals and used to anticipate risk, cost, and retirement needs.

Review and exit triggers

  • an unlisted item is discovered;
  • ownership is unclear or lapses;
  • reconciliation shows a gap;
  • an item changes provider, data, or authority; or
  • a review date passes.

3.3 Risk and consequence tiering

Full record: worked example 2 of 5 in Part 3, Capability model. The catalog entry in Part 4 gives the intended outcome, workloads, delivery modes, and condition.

3.4 Responsible-use and approval review

Purpose

Prevent consequential uses from going ahead without anyone examining their ethical, social, legal, and business effects.

Intended outcome

Proposed uses receive ethical, social, legal, and business review proportional to their consequences.

Decisions enabled

  • whether a use may proceed, proceed with conditions, or not proceed;
  • what conditions apply;
  • which concerns are accepted and by whom; and
  • when a use must return for review.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • define which uses need review and at what depth, using consequence tiers;
  • set criteria covering affected people, rights, fairness, transparency, and business effects;
  • include reviewers with relevant perspectives, including affected communities where appropriate;
  • record decisions, conditions, and unresolved concerns;
  • check that conditions are met before and after launch; and
  • revisit approvals when use, data, or authority changes.

Required evidence

  • review criteria and tiers;
  • decisions with reasoning;
  • conditions and closure evidence;
  • unresolved concerns with owners; and
  • re-review records.

Common failure patterns

  • review that approves what the sponsor already decided;
  • review depth that ignores consequence;
  • conditions never checked;
  • reviewers without independence or expertise; and
  • approval treated as permanent.

Capability levels

Level Observable condition
Emerging Review depends on who notices a concern, and decisions are not recorded.
Defined Defined criteria and a review path exist for in-scope uses, with recorded decisions.
Managed Depth scales with consequence, conditions are tracked to closure, and re-review occurs on change.
Adaptive Criteria and practice evolve with experience, and reviews draw on outcomes, complaints, and incidents from live operation.

Review and exit triggers

  • use, data, or authority changes materially;
  • a complaint, incident, or finding raises a new concern;
  • conditions are not met;
  • law or stakeholder expectations change; or
  • a scheduled re-review date.

3.5 Regulatory and contractual obligations

Purpose

Prevent legal, contractual, licensing, and professional obligations from being found after a deployment has already breached them.

Intended outcome

Applicable laws, regulations, contracts, licenses, and professional obligations are identified and maintained.

Decisions enabled

  • which obligations apply to a use;
  • whether a use is compliant or needs change;
  • what contract terms are required; and
  • when obligations must be re-examined.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • identify applicable laws, regulations, contracts, licenses, and professional duties by use, location, and affected people;
  • assess each use against them with legal and subject experts;
  • record obligations in a register with owners;
  • require contract terms that support the obligations;
  • monitor changes in obligations; and
  • record decisions about interpretation and residual risk.

Required evidence

  • obligation register;
  • assessments and interpretations;
  • required contract terms;
  • change monitoring records; and
  • residual-risk decisions.

Common failure patterns

  • relying on the vendor's statement of compliance;
  • obligations assessed once;
  • license terms for models and data ignored;
  • multi-jurisdiction use assessed under one law; and
  • obligations known to legal staff but not to builders.

Capability levels

Level Observable condition
Emerging Obligations are considered when someone raises them, and knowledge sits with a few individuals.
Defined A register of applicable obligations exists with owners and a review of each in-scope use.
Managed Obligations drive design, contracts, and evidence, and changes are monitored and acted on.
Adaptive The organization anticipates regulatory and contractual change and adjusts designs before obligations take effect.

Review and exit triggers

  • a new law, regulation, or standard takes effect;
  • use extends to a new location, population, or data type;
  • a contract or license changes;
  • a regulator or counterparty inquires; or
  • an assessment is found to be outdated.

3.6 Transparency, explanation, and recourse

Purpose

Prevent people affected by AI from being unaware of its involvement, unable to understand important limits, or left without a practical path to a human, correction, appeal, or remedy.

Intended outcome

People affected by AI receive appropriate notice, understandable information about AI involvement and important limits, meaningful choices, and accessible paths to human help, correction, appeal, or remedy.

Decisions enabled

  • when and how to notify people;
  • what to explain and at what level of detail;
  • what choices and alternatives to offer;
  • which recourse paths are required; and
  • whether the arrangements work for the people who need them.

Applicability

Workloads: All
Delivery modes: All
Condition: Where people are affected
Typical state: Conditional

Key activities

  • identify who is affected and in what way;
  • set a disclosure standard covering notice, explanation, and important limits;
  • provide choices, including a non-AI route where appropriate;
  • design accessible routes to human help, correction, appeal, and remedy;
  • test notices and recourse with affected people, including those with access needs; and
  • track use and outcomes of recourse and fix barriers.

Required evidence

  • disclosure standard;
  • notices and explanations as delivered;
  • recourse paths and responsible roles;
  • accessibility and effectiveness tests; and
  • recourse volumes, timeliness, and outcomes.

Common failure patterns

  • disclosure buried in terms;
  • explanations that are accurate but unusable;
  • an appeal path that leads back to the same system;
  • recourse that exists but takes too long to help; and
  • no test with affected people.

Capability levels

Level Observable condition
Emerging Notice and help are provided case by case, and affected people often cannot tell AI was involved.
Defined A standard defines notice, explanation, and recourse for in-scope uses, and paths are published.
Managed Notices and recourse are tested with affected people. Outcomes and delays are tracked and barriers removed.
Adaptive Practice adapts to complaints, outcomes, and changes in what people expect, and recourse results inform design and approval of uses.

Review and exit triggers

  • a use begins to affect people not previously considered;
  • complaints or recourse volume rise;
  • law or guidance on disclosure changes;
  • a test shows people cannot use the paths; or
  • the AI's behavior or limits change.

3.7 Fairness and disparate-impact evaluation

Purpose

Prevent materially different outcomes for different groups from going unnoticed or unaddressed because only average performance is checked.

Intended outcome

The organization evaluates and responds to materially different outcomes across affected groups.

Decisions enabled

  • whether outcomes differ enough to require action;
  • what corrective action to take, or whether to stop;
  • which groups and measures matter; and
  • how often to repeat the evaluation.

Applicability

Workloads: All
Delivery modes: All
Condition: Where outcomes can differ materially across affected groups
Typical state: Conditional

Key activities

  • decide whether the use creates outcomes that can differ materially across affected groups;
  • define groups, measures, and thresholds with legal, domain, and affected-community input;
  • evaluate before release and at intervals during operation;
  • investigate causes, including data, design, and process;
  • apply and verify corrective actions; and
  • record limits, such as groups that cannot be measured.

Required evidence

  • assessment method and rationale for scope;
  • results by group;
  • thresholds and decisions;
  • corrective actions and verification; and
  • documented limitations.

Common failure patterns

  • testing only overall accuracy;
  • groups defined after results are seen;
  • no data to assess groups, with no alternative method;
  • findings recorded and not acted on; and
  • evaluation done once.

Capability levels

Level Observable condition
Emerging Outcome differences are examined only after complaints or by individual initiative.
Defined In-scope uses have a defined method, groups, and thresholds, and results are recorded.
Managed Evaluation repeats on a schedule and after change. Corrective actions are verified and decisions are recorded.
Adaptive Methods improve through experience and external practice, and evaluation informs design choices, data decisions, and approval of uses.

Review and exit triggers

  • results exceed a threshold;
  • data, model, or population changes;
  • a complaint or external report indicates unequal outcomes;
  • the use extends to new groups; or
  • a scheduled evaluation date.

3.8 Human oversight and escalation

Purpose

Prevent human oversight from becoming nominal, where the person nominally in charge cannot see enough, act in time, or override the AI.

Intended outcome

Human authority is meaningful, timely, competent, and able to interrupt, reverse, or escalate AI-supported work.

Decisions enabled

  • where human authority is required and what form it takes;
  • who exercises it and with what training;
  • when AI may act without prior approval; and
  • when oversight must be strengthened or authority reduced.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • define what the overseer must see, decide, and be able to do;
  • set response times, staffing, and competence requirements;
  • provide the means to interrupt, reverse, or escalate;
  • protect overseers from pressure to approve without review;
  • test whether oversight catches errors, including ones that appear plausible; and
  • record interventions and use them to adjust design.

Required evidence

  • oversight design and authority;
  • staffing and training;
  • intervention and override records;
  • effectiveness tests; and
  • escalation records.

Common failure patterns

  • approval volume exceeds capacity to review;
  • overseers lacking information or authority;
  • no ability to reverse effects;
  • oversight that exists only on the organization chart; and
  • override rates never examined.

Capability levels

Level Observable condition
Emerging Oversight depends on individual diligence, and authority and capability are not defined.
Defined Oversight roles, authority, and means to intervene are defined for in-scope uses.
Managed Staffing and competence match the load. Effectiveness tests and intervention records lead to changes.
Adaptive Oversight adapts to observed error, volume, and autonomy, and leaders use test results to decide where to extend or reduce AI authority.

Review and exit triggers

  • intervention or error rates change;
  • volume exceeds overseer capacity;
  • AI authority or tools expand;
  • an incident shows delayed or missing intervention; or
  • a test shows overseers miss seeded errors.

3.9 Audit, assurance, and reporting

Purpose

Prevent leaders and reviewers from relying on assurances they cannot check, by ensuring that reliable evidence about design, operation, outcomes, incidents, and obligations can be obtained.

Intended outcome

Decision makers and reviewers can obtain reliable evidence about design, operation, outcomes, incidents, and obligations.

Decisions enabled

  • whether controls and outcomes are as represented;
  • what evidence is needed and how often;
  • which findings need action; and
  • what is reported to leaders, regulators, and counterparties.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • define the evidence required for each type of claim or obligation;
  • keep an evidence register with source, owner, and retention;
  • perform assurance reviews proportionate to consequence, using independent reviewers where needed;
  • report findings to decision makers in a form they can act on;
  • track findings to closure; and
  • protect the evidence from alteration.

Required evidence

  • evidence register;
  • assurance plans and results;
  • management and board reporting;
  • findings and closure; and
  • protections on evidence integrity.

Common failure patterns

  • evidence assembled only when an audit is announced;
  • reviewers who rely on documents without testing operation;
  • reports of activity rather than conditions;
  • findings that remain open; and
  • independent review missing for high-consequence uses.

Capability levels

Level Observable condition
Emerging Evidence is collected on request and varies in quality.
Defined An evidence register and a review schedule exist for in-scope uses.
Managed Assurance is proportionate and sometimes independent. Reports reach decision makers and findings are closed.
Adaptive Evidence is gathered continuously where practical, assurance adapts to risk and incident patterns, and leaders use it to change decisions.

Review and exit triggers

  • a finding stays open past its date;
  • evidence cannot be produced on request;
  • a regulator or counterparty requires new evidence;
  • an incident reveals unreported conditions; or
  • a use becomes more consequential.

3.10 Unapproved-AI discovery and onboarding

Purpose

Prevent unapproved AI use from staying hidden, where data may leave control and no one can assess, support, or stop it.

Intended outcome

The organization discovers unapproved use and provides a practical path to assess, onboard, constrain, replace, or stop it.

Decisions enabled

  • how to treat each discovered use: assess, onboard, constrain, replace, or stop;
  • where to improve approved options to reduce the need for workarounds;
  • which signals to monitor; and
  • how quickly responses must occur.

Applicability

Workloads: All
Delivery modes: B, C
Typical state: Conditional, when unapproved use is found or likely

Key activities

  • look for unapproved use through purchasing, network, identity, and survey signals within what law and policy permit;
  • ask people why they use the tool and what need it meets;
  • assess each item using the same tiering and review as approved uses;
  • provide a fast path to approve, constrain, or replace;
  • stop or contain uses that cannot be accepted, and explain why; and
  • track time to disposition and recurring causes.

Required evidence

  • discovery results and methods;
  • disposition records;
  • onboarding time;
  • reasons for use; and
  • changes made to approved options.

Common failure patterns

  • punishing disclosure, which hides more use;
  • discovery without an onboarding route;
  • ignoring the unmet need;
  • treating all unapproved use as equally risky; and
  • one-time surveys.

Capability levels

Level Observable condition
Emerging Unapproved use is found by accident and handled by blocking or ignoring it.
Defined Discovery methods and a path for assessing discovered uses exist.
Managed Dispositions are timely and consistent. Reasons for use lead to better approved options.
Adaptive Discovery adapts to new tools and behavior, and the organization reduces unapproved use by meeting needs, not by increasing enforcement alone.

Review and exit triggers

  • discovery finds sensitive data in an unapproved tool;
  • onboarding time exceeds the agreed target;
  • the same need recurs across teams;
  • a new class of tool appears; or
  • the approved options change.